Research map / Architectures and efficiency / RNNs and normalization
DynaNav: Dynamic Feature and Layer Selection for Efficient Visual Navigation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DynaNav, a visual navigation model that combines a trainable Gumbel-Softmax hard feature selector with early-exit decoder layers whose thresholds are tuned by Bayesian optimization under FLOPs, time and memory constraints.
Problem
Transformer-decoder visual navigation foundation models (e.g., ViNT, NoMaD) are computationally expensive and opaque, which hinders deployment on edge devices.
Approach
Combine a pixel-wise trainable hard feature selector (Gumbel-Softmax masks) with an early-exit mechanism over decoder layers, where exit thresholds are chosen post-training by Bayesian optimization subject to FLOPs, inference-time and GPU-memory constraints, so that compute adapts to scene complexity.
Key results
- Versus ViNT, DynaNav reduces FLOPs from 4.37e9 to 1.93e9 on RECON (2.26x) with 42.3% lower inference time and 32.8% lower memory (13.35 GB vs 19.07 GB).
- On the four benchmarks DynaNav improves Sim(a_t) and Sim(w_t) slightly over ViNT on average (e.g., RECON action similarity 94.92 vs 94.49), while NoMaD keeps a ~0.2% higher average waypoint performance at about 4x FLOPs.
- In CARLA Scene A, DynaNav reaches a success rate of 0.727 versus 0.724 for ViNT with 1.58e9 vs 4.37e9 FLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DynaNav instead of ViNT | visual navigation model · Recon, Go-Stanford, SACSoN, and SCAND | action cosine similarity: 0.83% improvement | Helped |
| use DynaNav instead of ViNT | visual navigation model · Recon, Go-Stanford, SACSoN, and SCAND | FLOPs: about 58% FLOPs | Helped |
| replace Transformer decoder blocks with Mamba blocks | DynaNav · RECON | waypoint cosine similarity: 95.09 vs 96.53 | Hurt |
| replace dynamic decoder with half the decoder layers (ablation) | DynaNav · Recon | action cosine similarity: 94.92 -> 91.05 | Hurt |
| reduce hidden channel size from C to C/2 (ablation) | DynaNav · Recon | action cosine similarity: 94.92 -> 89.70 | Hurt |
| remove dynamic decoder (ablation) | DynaNav · Recon | action cosine similarity: 94.92 -> 94.49 | Hurt |
| add feature selector (ablation) | DynaNav without feature selector · Recon | action cosine similarity: 93.68 -> 94.92 | Helped |
| add feature selector (ablation) | DynaNav without feature selector · Recon | FLOPs ($10^9$): 4.06 -> 1.93 | Helped |
| remove post-training Bayesian Optimization (ablation) | DynaNav · RECON | waypoint cosine similarity: 96.58 -> 96.30 | Hurt |
| allow early exit before the decoder (ablation) | DynaNav with BO · RECON | successful rate: 0.732 -> 0.727 | Hurt |
| remove FLOPs constraint (ablation) | DynaNav threshold optimization · RECON | FLOPs ($10^9$): 1.93 -> 2.84 | Hurt |
| remove time constraint (ablation) | DynaNav threshold optimization · RECON | Time (s/traj): 0.228 -> 0.273 | Hurt |
Surprising findings
- Without Bayesian-optimized thresholds, early exit degrades performance, and allowing exit before the decoder trades a small accuracy loss for efficiency.
- Simply halving decoder layers or channels reduces compute but gives poor generalization, with notably lower test accuracy than the full model.
Limitations
- Requires an additional Bayesian optimization stage after training, which adds labor cost.
- Efficiency gains vary by scene: FLOPs increase on harder outdoor and CARLA scenes with larger observation-goal discrepancy.
- Evaluated mainly on offline benchmarks and a single CARLA setup; code and weights not released.
Experimental setup
Datasets: RECON Go-Stanford SACSoN SCAND CARLA Town01 trajectories
Benchmarks: RECON Go-Stanford SACSoN SCAND CARLA Town02/Town03/Town10
Metrics: Sim(a_t, a_gt) Sim(w_t, w_gt) L_action L_dist FLOPs inference time GPU memory CARLA success rate
Baselines: ViNT NoMaD GNM
Builds on: ViNT DeeR-VLA Gumbel-Softmax EfficientNet
Closest papers by meaning
- Dynam3D: Dynamic Layered 3D Tokens Empower VLM for Vision-and-Language Navigation (NeurIPS 2025 · similarity 0.88)
- Seeing through Uncertainty: Robust Task-Oriented Optimization in Visual Navigation (NeurIPS 2025 · similarity 0.86)
- ImagineNav: Prompting Vision-Language Models as Embodied Navigator through Scene Imagination (ICLR 2025 · similarity 0.86)
- MemoNav: Working Memory Model for Visual Navigation (CVPR 2024 · similarity 0.86)
- Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator (NeurIPS 2026 · similarity 0.86)
- NaviFormer: A Spatio-Temporal Context-Aware Transformer for Object Navigation (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.